Mistral AI, in partnership with All Hand Bress AI, has released an updated version of a large developer-centric language model Devstral 2507 Label. This version includes two models –Open air small 1.1 and Devstral Medium 2507– Designed to support agent-based code inference, program synthesis, and structured task execution for large software repositories. These models are optimized for performance and cost to suit real-world use in developer tools and code automation systems.
Devstral Small 1.1: Open models for local and embedded use
Open air small 1.1 (also known as devstral-small-2507) Based on the Mistral-Small-3.1 basic model, it contains about 24 billion parameters. It supports a 128K token context window that allows it to handle multi-file code input and long prompts typical of software engineering workflows.
The model is specifically designed for structured output, including XML and functional title formats. This makes it compatible with proxy frameworks such as OpenHands, and is suitable for tasks such as program navigation, multi-step editing, and code search. It is licensed under Apache 2.0 and can be used for research and commercial purposes.
Performance: SWE basic results
Devstral Small 1.1 Achievements 53.6% In a benchmark tested by SWE Bench, this benchmark evaluates the model’s ability to generate the correct patches for actual GitHub issues. This represents a significant improvement to the previous version (1.0) and put it ahead of other publicly available models of comparable sizes. The results were obtained using OpenHands scaffolding, which provides a standard testing environment for evaluating code agents.
Although not the largest proprietary model level, this version provides a balance between size, inference cost, and inference performance, which is practical for many coding tasks.
Deployment: Local reasoning and quantization
The model is published in a variety of formats. Quantitative versions in GGUF are available for llama.cpp,,,,, vLLMand LM Studio. These formats enable the operation of reasoning locally on high memory GPUs (such as the RTX 4090) or Apple Silicon machines with 32GB RAM or more. This is beneficial for developers or teams who do not rely on managed APIs and are unwilling to operate.
Mistral can also provide the model through its inference API. The current pricing is $0.10 per million input tokens and $0.30 per million output tokens, the same as other models in the Mistral-Small Line.
Devstral Medium 2507: Higher precision, API only
Devstral Medium 2507 Not open source, only available through the Mistral API or enterprise deployment protocol. It offers the same 128K token context length as the smaller version, but with higher performance.
Model score 61.6% Based on the proven SWE, several business models including Gemini 2.5 Pro and GPT-4.1 are included in the same evaluation framework. It has a stronger inference ability in long contexts, making it a candidate for code agents that run in large monorepos or repositories with cross-file dependencies.
API pricing is set to $0.40 per million USD input token and $2 per million USD output tokens. The Mistral platform allows enterprise users to fine-tune.
Comparison and use case fitting
| Model | SWE bench verified | Open source | Enter cost | Output cost | Context length |
|---|---|---|---|---|---|
| Open air small 1.1 | 53.6% | Yes | $ 0.10/m | $0.30/m | 128K token |
| Prejudice medium | 61.6% | No | $0.40/m | $2.00/m | 128K token |
Devstral Small is more suitable for local development, experimentation or integration into client developer tools where control and efficiency are important. By contrast, Devstral media provides greater accuracy and consistency in structured code editing tasks and is designed to benefit from higher performance despite cost increases.
Integrate with tools and agents
Both models are designed to support integration with code proxy frameworks such as OpenHands. Support for structured function calls and XML output formats enables them to be integrated into automated workflows to generate tests, refactors, and bug fixes. This compatibility makes it easier to connect Devstral models to IDE plugins, version control robots, and internal CI/CD pipelines.
For example, developers can use small developers to prototypify local workflows, while Devstral media can be used in production services to apply patches or Triage drag requests based on model suggestions.
in conclusion
Devstral 2507 version reflects targeted updates to Mistral’s code LLM stack, providing users with a clearer tradeoff between inference cost and task accuracy. Devstral Small provides an accessible open model for many use cases with sufficient performance, while Devstral Medium is suitable for applications where correctness and reliability are critical.
The availability of the two models under different deployment options makes them relevant between stages of the software engineering workflow – from experimental agent development to deployment in a business environment.
Check Technical details,,,,, Prejudice when hugging faces small models Devstral Media will also be available on Mistral code for enterprise customers and the Fineting API. All credits for this study are to the researchers on the project. Also, please feel free to follow us twitterand Youtube And don’t forget to join us 100K+ ml reddit And subscribe Our newsletter.
Sana Hassan, a consulting intern at Marktechpost and a dual-degree student at IIT Madras, is passionate about applying technology and AI to address real-world challenges. He is very interested in solving practical problems, and he brings a new perspective to the intersection of AI and real-life solutions.
